Skip to content
Open access

Eye-Preserving 3D Face Registration Using Single-Image Reconstruction and Failure-Type Stratification

2026 · Computer Modeling in Engineering & Sciences · 0 citations · 26 references

Abstract

: Registering a textured 3D face mesh reconstructed from one image to a high-quality 3D face scan is challenging because the inputs come from different pipelines. Even for the same subject, they may differ in scale, pose convention, facial extent, expression, eye state, and artifacts. Minimizing nearest-neighbor surface distance alone can yield geometrically implausible alignments or deform important regions such as the eyes. We propose a branch-stratified registration-and-recovery pipeline for aligning open-eye reconstructed source meshes to eye-closed target 3D face scans. The method combines orientation-gated multi-start rigid registration, pre-refinement failure-type stratification, and eye-preserving non-rigid refinement with an eight-stage nasal depth-contour schedule within 16 scheduled updates (S8). Every processed output is assessed by the same post-S8 quality-control (QC) conjunction, which requires valid direct target-anchor evidence, facial orientation, full-surface and nasal errors, semantic nose-anchor consistency, edge strain, and a fixed-eye implementation audit. A deterministic identity-disjoint split assigned 10 FaceScape subjects to development and 10 to revision-stage evaluation, with 19 expressions nested within each subject. On the frozen Hierarchical Representation Network (HRN) evaluation partition, the proposed method accepted 168/190 pairs, corresponding to a mean subject acceptance rate of 88.42% (whole-subject bootstrap 95% confidence interval (CI), 67.89%–100.00%). Under the same target and QC protocol, 3D Dense Face Alignment version 2 (3DDFA-V2) achieved 164/190 (86.32%), whereas the shared-cues Open3D rigid control achieved 101/190 (53.16%). Across all 190 S8 outputs in the HRN evaluation partition, the mean subject-level full-surface and nasal median distances were 2.57 mm (95% CI, 2.05–3.27) and 0.90 mm (0.76–1.04), respectively. A development-selected full-resolution, eye-constrained as-rigid-as-possible (ARAP) baseline completed all 190 attempts. Compared with S8, ARAP produced larger full-surface (3.23 vs. 2.57 mm) and nasal (2.41 vs. 0.90 mm) median distances but lower p99 edge strain (0.046 vs. 0.162) and p90 displacement (0.49 vs. 2.47 mm), revealing a fit–deformation tradeoff rather than uniform superiority of either method. In a technical assessment masked to case metadata and automatic QC results, two non-author raters evaluated every HRN output. A third non-author rater independently assessed the five cases on which they disagreed without seeing either primary rating; two-of-three majority labels were then used for those cases. The resulting visual reference classified 169/190 outputs as usable; against this reference, the frozen QC achieved 100% positive predictive value, 99.41% sensitivity, and 100% specificity.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.